acceptodds
Under review as a conference paper at ICLR 2027

When Local Classifiers Become Representations: The Plasticity Tax of Unsupported Outputs

Abstract

In class-incremental local learning, classifier outputs can become downstream representations before all classes are observed. Outputs for classes without positive examples remain unsupported, yet participate in local softmax competition and receive updates. We investigate the representation-level plasticity tax through matched interventions on unsupported-output count, competition strength, exposure depth, and training phase, holding capacity, training data, replay, and optimization budgets fixed. Local softmax analysis motivates Anchor-Forward, an exact intervention on the unsupported-coordinate update path. At a common input and model state, it preserves forward computation and supported-logit gradients while blocking direct local gradients through unsupported coordinates. In DeeperForward, depth and phase interventions identify the largest losses in shallow layers and during new-task acquisition. On CIFAR-100, Anchor-Forward improves arriving-task accuracy by 27.66 percentage points over matched competition in a two-task experiment with 80 unsupported outputs. In a ten-task stream, it recovers 16.13 points in mean arriving-task accuracy. Experiments with HCL-FF and GTSRB demonstrate plasticity recovery across architectures and domains. These findings indicate that output support is part of representation design when local classifier responses are propagated to subsequent learners.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Class-Incremental Learning, Continual Plasticity, Unsupported Outputs, Output Support, Propagated Local Classifiers, Gradient Support

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_52byOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines an important aspect of class-incremental learning: not only retaining previously learned knowledge, but also maintaining the ability to learn new classes. The authors investigate how unsupported outputs (classes not yet seen during training), can negatively affect learning when local classifier outputs are used as representations for subsequent layers. They show that these outputs participate in local softmax competition and receive gradient updates despite having no positive training examples, potentially impairing the acquisition of new classes. To address this issue, the paper proposes Anchor-Forward, a method that preserves forward competition while blocking direct gradients through unsupported outputs. The authors further analyze how the number of unsupported outputs, competition strength, layer depth, and exposure timing influence learning performance. Experiments on CIFAR-100 and GTSRB using DeeperForward and HCL-FF show that the proposed intervention can substantially recover new-task learning performance.

Critical strengths

The paper addresses an important yet relatively underexplored aspect of continual learning by focusing on the ability to acquire new knowledge rather than solely on preventing forgetting. The proposed Anchor-Forward intervention is simple yet effective, showing substantial improvements in new-task accuracy across different architectures and datasets.

Critical weaknesses
  1. The negative effects of unsupported outputs could be easily avoided by incrementally expanding the classifiers or simply masking unseen classes during training. This raises concerns about whether the problem reflects realistic learning scenarios or is largely an artifact of the experimental setup. As a result, the paper seems to artificially introduce an otherwise avoidable problem and then address it with an unnecessarily complicated intervention, raising fundamental questions about the practical relevance and necessity of the proposed method.

  2. The experiments assume a predefined output capacity, which requires knowing the total number of classes in advance. It would be much more practical to consider dynamic classifier expansion, where the output layer grows as new classes arrive.

  3. The proposed Anchor-Forward consistently performs worse than the simpler M0 baseline in new-task accuracy. This raises questions about its practical value beyond serving as an analytical intervention, as the justification for using the proposed method over a simpler and better-performing alternative remains unclear.

  4. The provided supplementary code appears to implement a substantially different method from the one described in the paper. Specifically, it does not include the proposed stop-gradient mechanism or the unsupported-class setting, making it difficult to reproduce the main experiments.

  5. The evaluation is limited to relatively small datasets and architectures, and even CIFAR-100 is evaluated using only a subset of its training samples. Experiments with larger datasets and more realistic training settings would strengthen the findings.

  6. Although the effect may be more pronounced in FF architectures that propagate intermediate classifier outputs, unsupported-output gradients are not inherently specific to FF. The appendix includes conventional backpropagation experiments, but these are too limited to establish whether the findings generalize to broader learning settings.

  7. The stronger negative effects observed in shallow layers may depend on the specific architecture and learning mechanism of FF models. It remains unclear whether the same behavior would occur in networks trained with standard backpropagation.

  8. The paper lacks sufficient conceptual discussion and experimental comparisons with prior work. Existing studies have reported similar issues with negative supervision [1] and even contrasting findings on the benefits of pre-allocated future-class outputs [2], but these connections are not adequately discussed. More concrete comparisons with related methods like [3–5] are needed to establish the novelty and generality of the proposed approach.

[1] Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning

[2] Class-incremental Learning with Pre-allocated Fixed Classifiers

[3] A Contrastive Symmetric Forward-Forward Algorithm for Continual Learning Tasks

[4] SS-IL: Separated Softmax for Incremental Learning

[5] Loss of Plasticity in Deep Continual Learning

Questions for authors

The authors should address the concerns raised above, particularly by clarifying the practical relevance and justification of the proposed experimental setting and method. Given that simpler alternatives can avoid the identified problem, it is important to explain why Anchor-Forward is necessary and under what conditions it offers meaningful benefits. Furthermore, broader evaluations using standard backpropagation and more diverse tasks and architectures are essential to establish the generalizability and significance of the findings.

6 comments
AuthorsOpenReview ↗

Response 1: W1, W2 — Practical relevance and class allocation

Opening

Thank you for your detailed review and for identifying where stronger evidence and clearer documentation were needed. Your comments motivated experiments using the full training pool and further controls on the training configuration. We also revisited the research code and clarified the implementation details. Below, we respond to your comments individually and explain the corresponding revisions.

Added theory and experiments

The full-pool study uses 48,000 gradient-training and 2,000 validation images. Terminal and native-auxiliary BP controls and a common-prefix study test the limits of the local result (Sections 5.6–5.7; Appendix I). We separate the depth cohorts rather than pooling different exposure histories (Section 5.2; Appendix D.2–D.3). A fixed 20-continuation follow-up from five native-local prefixes adds a positive conditional cut interaction.

We have updated the supporting materials with the complete code for our method. Experiment configurations and result-recomputation scripts are also included.

Figure changes

Figure 3(b) and Figure S5 separate the depth and phase cohorts. Figure S10 removes the pooled heterogeneous depth estimate and uses the original three-seed current-versus-neither comparison. Added Figure S11 displays five prespecified paired contrasts from the full-pool, auxiliary-BP, and common-prefix studies, with their distinct endpoints and intervals. Figure 7 in main-text Section 5.7 shows the new native-local-prefix seed pairs and its two contrasts prespecified in this follow-up.

W1 | Avoidability and practical relevance

Unsupported outputs can be masked. The experiment asks a narrower question: can acquisition recover while numerical denominator participation remains? The common-backbone local comparison gives 17.10/43.60/44.72% current accuracy for Matched/Anchor/M0, and the full-pool comparison gives an 18.451-point arriving gain over Matched. These results identify acquisition recovery while retaining denominator participation.

The architectural motivation is that class-associated feature maps serve two roles: they are locally supervised, and they become inputs to later blocks. Their forward role persists even when the corresponding class has not yet supplied positive examples. This makes the interaction between local supervision and later representation learning a concrete question. Our comparison keeps the feature bank and its forward transmission, and changes how unsupported coordinates participate in the local loss. It shows an avoidable acquisition cost in the tested configuration and a way to isolate its derivative contribution.

Changes made: Stated the diagnostic question tested by Anchor (Introduction; Section 6).


W2 | Unknown class count

Our experiments assume a known preallocated output capacity. Expanding class-associated feature groups can also change the next block's input width; masking a fixed bank is not dynamic expansion. The full-pool study still assumes 100-class capacity.

Changes made: Specified the known-capacity assumption and distinguished fixed-bank masking from dynamic expansion (Section 2; Section 6).


AuthorsOpenReview ↗

Response 2: W3, W4 — M0 comparison and updated complete code

W3 | Comparison with M0

M0 has higher observed new-task means in all three comparisons. Original two-task current accuracy is 44.84% for Anchor and 46.40% for M0. The common-backbone two-task values are 43.60% and 44.72%. Full-pool mean arriving accuracy is 71.033% and 71.858%, respectively. These comparisons measure acquisition separately from final accuracy. Anchor tests whether acquisition can recover while unsupported coordinates remain in numerical normalization. Relative to Matched, the original and full-pool gains are 27.66 and 18.451 percentage points. Removing direct unsupported derivatives therefore recovers much of the acquisition loss while retaining denominator coordinates.

Changes made: Compared the acquisition means of Anchor and M0 directly (Sections 5.3, 5.6–5.7 and 6; Appendix I.8).


W4 | Code and reproducibility

The updated supporting materials contain complete source code for our method implementations in DeeperForward and HCL-FF. In each implementation, M0 restricts local normalization to supported classes, Matched includes the selected unsupported classes, and Anchor detaches only their loss-side scores. Class-associated feature maps still propagate. DeeperForward uses native sequential local updates; HCL-FF retains its local contrastive objective, projection heads, routed cross-entropy, and causal terminal classifier. The README identifies these distinct call chains and their readouts.

The implementation distinction is explicit in the native call chains. DeeperForward computes local cross-entropy from pooled goodness, but sends the class-associated feature maps to the next block. Anchor detaches the selected unsupported scores in the loss calculation; it does not detach or replace those propagated maps. HCL-FF applies the intervention through its routed cross-entropy while retaining its local contrastive components. The three branches therefore exercise the claimed unsupported-class setting and loss-side stop-gradient in the actual research architectures. The source manifest records the files used by these adapters, so the branch definitions can be checked against the native calls instead of inferred from a method name.

From a clean extraction, both full-width 17-layer, 100-class implementations completed all three branches for three epochs of two batches on eight real CIFAR-100 training examples. Every local optimizer completed six updates; states remained finite, and saved model checkpoints reproduced native predictions exactly after reloading. The command is python native_three_arm_small_training.py --kind DF --data-root DATA_ROOT --out DF_small.json --checkpoint-dir LOCAL_SMOKE_CHECKPOINTS; replace DF with HCL for the second implementation.

The updated package, 39008_Reviewer_Code_Materials.zip, also includes experiment configurations, run instructions, saved results, and result-recomputation scripts. Its README explains the entry points, and the reviewer guide links each concern to the relevant implementation and records.

Changes made: Updated the supporting materials with complete DF/HCL method implementations, experiment configurations, run instructions, and result-recomputation scripts.


AuthorsOpenReview ↗

Response 3: W5, W6 — Full-pool experiments and BP controls

W5 | Full training pool and larger settings

The added CIFAR-100 study uses all 50,000 official training images, split into 48,000 gradient-training and 2,000 validation images. Four systems and five seeds cover ten tasks; all 190 training units precede unified test evaluation. Anchor-minus-Matched mean arriving accuracy is +18.451 pp, while Anchor-minus-M0 final accuracy is -0.072 pp. These are the two prespecified paired comparisons; their 97.5% intervals are reported in Table 19. Final old-class accuracy remains below 1% for all systems. Tiny-ImageNet is mixed: the preset seed-0 gain is +4.8 points, below the preset 5-point criterion. Food101 has no Anchor arm.

This is a complete training-pool check, with validation separated from gradient updates. Each of the 100 classes contributes 480 training and 20 validation images. The same four systems follow ten ten-class tasks, using the recipes selected in the earlier two-task comparison without retuning on the full-pool test results. At each task boundary, prediction takes argmax over all seen classes: “current” and “old” identify subsets of test examples, not a smaller allowed label set. For each seed, arriving accuracy first averages tasks 2–10; paired uncertainty then uses the five seed-level differences. The 190 units comprise shared first-task prefixes and subsequent task training, yielding 20 system-by-seed trajectories, not 190 independent replications. This preserves the paired acquisition comparison while making its data coverage and statistical unit explicit.

Table R3. Full-pool CIFAR-100 results

SystemArriving (%)Final (%)Old 90 (%)
Local M071.858 ± 0.2538.502 ± 0.1750.718
Local Matched52.582 ± 1.4048.484 ± 0.3990.927
Local Anchor71.033 ± 0.6668.430 ± 0.2990.738
Terminal BP-M079.027 ± 0.6128.984 ± 0.1880.758

48,000 training and 2,000 validation images; ten tasks and five seeds. Arriving accuracy averages tasks 2–10. Arriving and final values are mean ± sample SD; Old 90 is the final mean. Sources: Section 5.6, Table 3 and Appendix Table 19.

Changes made: Added the full-pool study and paired results, and reported the exploratory Tiny-ImageNet and Food101 scope (Section 5.6; Appendix I.3, I.6–I.7; Table 19; Figure S11).


W6 | Beyond local learning

We added terminal-only and native-auxiliary-loss BP interventions. Current Anchor-minus-Matched effects are -0.500 and -0.520 pp, respectively. The full-pool terminal BP-M0 baseline also has higher arriving/final means than local Anchor. The common-prefix cut experiment then controls the starting state, recipe, losses, and readout; its interaction is -2.280 pp.

We separate a cross-system comparison from a controlled change in gradient connectivity. Terminal BP uses a final classifier; native-auxiliary BP sums all 17 local losses and predicts from the 17 native scores. In the cut tests, the common engine sums those losses, performs one backward pass and steps all 17 optimizers in both Cut and Coupled. Cutting block-input gradients leaves the forward feature maps and residual paths present. The 2×2 comparison then asks whether this backward change alters the Anchor-minus-Matched effect.

The new follow-up restores five native-local M0 first-task states and runs 20 fixed continuations in one common summed-loss engine. Cut and Coupled differ only in block-input detachment, with optimizer moments, replay and input/dropout streams preserved. Both prefix studies use the same constant-rate second-task recipe, which differs from the original native sequential-update experiment. Current accuracy is 26.46/37.16% for Cut Matched/Anchor and 30.92/27.52% for Coupled Matched/Anchor, always predicting among all 20 seen classes. The Cut effect is +10.700 pp; the Cut-minus-Coupled interaction is +14.100 pp. All 20 final states were locked before testing. Backward cuts strengthen the relative Anchor acquisition effect under this native-local starting configuration and continuation recipe. All four arms and old/current/final/AIA endpoints are reported.

Changes made: Added the native-local-prefix control while retaining the terminal BP, auxiliary BP, and auxiliary-prefix results (Section 5.7; Appendix I.2, I.4–I.5 and I.8; Figures 7 and S11).


Reviewer_N2MYOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the loss of plasticity in forward-forward networks. Specifically, it focuses on unsupported outputs, which are logits that are not used in current tasks but receive gradient signals in class-incremental settings. The paper argues that leaving those outputs remain in softmax operation creates "plasticity tax", which degrades plasticity of the network. The paper proposes to include those outputs in softmax but stop gradient flow to maintain plasticity. The experimental results reveal correlation between number of unsupported outputs and plasticity, and the anchor-forward approach successfully maintains plasticity compared to naive training.

Critical strengths

None

Critical weaknesses
  1. Limited novelty
  • Unsupported output problem in class-incremental learning is already addressed in related works and often can be mitigated with simple interventions (e.g. excluding from softmax) [3-5]. It is hard to find what novelty exists in this paper in this context.
  1. Justification on the proposed method is weak
  • Simply excluding unsupported outputs (M0) improves plasticity and often better than the propose method. The only difference between M0 and the proposed method is that the proposed method includes unsupported outputs to softmax but there is no need for such design.
  1. The paper lacks discussion on related works and existing baselines
  • Loss of plasticity is largely studied in neural networks trained with backpropagation [1-2]. The paper does not include discussion on these related works.
  • Regarding the unsupported output problem in class-incremental learning, there has been related approaches [3-5]. The paper should discuss those approaches and include as baseline.

[1] Lyle, Clare, et al. "Understanding plasticity in neural networks." International conference on machine learning. PMLR, 2023.

[2] Dohare, Shibhansh, et al. "Loss of plasticity in deep continual learning." Nature 632.8026 (2024): 768-774.

[3] Caccia, Lucas, et al. "New insights on reducing abrupt representation change in online continual learning." arXiv preprint arXiv:2104.05025 (2021).

[4] Zeno, Chen, et al. "Task-agnostic continual learning using online variational bayes with fixed-point updates." Neural Computation 33.11 (2021): 3139-3177.

[5] Ahn, Hongjoon, et al. "Ss-il: Separated softmax for incremental learning." 2021 IEEE/CVF international conference on computer vision (ICCV). IEEE, 2021.

Questions for authors

See weakness

2 comments
AuthorsOpenReview ↗

Response 1: W1 — Contribution and comparison with masking

Opening

Thank you for carefully reviewing our work. We appreciate your questions about the contribution, its relationship to masking, and its place among existing methods. These questions prompted us to sharpen the research question and explain more clearly what the controlled experiments establish. We address each of your points below.

Added theory and experiments

The original same-state identity is retained, with a more precise account of its diagnostic purpose. Full-pool CIFAR-100 and BP controls add empirical tests (Sections 5.6–5.7; Appendix I). Section 7 distinguishes plasticity research and continual-classifier methods from our derivative controls. A fixed 20-continuation follow-up from five native-local prefixes adds a positive conditional cut interaction.

We have updated the supporting materials with the complete code for our method. Experiment configurations and result-recomputation scripts are also included.

Figure changes

Figure 3(b) and Figure S5 separate the depth and phase cohorts. Figure S10 removes the pooled heterogeneous depth estimate and uses the original three-seed current-versus-neither comparison. Added Figure S11 displays five prespecified paired contrasts from the full-pool, auxiliary-BP, and common-prefix studies, with their distinct endpoints and intervals. Figure 7 in main-text Section 5.7 shows the new native-local-prefix seed pairs and its two contrasts prespecified in this follow-up.

W1 | What is new relative to masking?

Our contribution is the controlled separation of normalization participation from direct local derivatives. Masking and unsupported-output supervision are established topics. The same-state identity appears in submitted Section 3; the revision clarifies its diagnostic purpose and adds empirical tests of its scope. M0 changes both; Anchor retains Matched's numerical loss and supported-logit derivatives at an identical state, while setting direct unsupported-logit derivatives to zero. In the original five-seed two-task study, current accuracy is 17.18/44.84/46.40% for Matched/Anchor/M0. The full-pool study supplies an additional 18.451-point Anchor-over-Matched arriving gain. These observations show that recovery can occur without excluding denominator coordinates.

The distinction is clearest at a shared model state. Matched normalizes over supported and unsupported scores, so an unsupported score contributes both a denominator term and a positive local loss derivative. M0 removes that score from normalization, changing the supported-score derivatives as well. Anchor keeps its numerical contribution but applies stop-gradient to that loss-side score. Consequently, the loss value and every supported-score derivative equal Matched's at that state, while the direct unsupported-score derivative is zero. This comparison separates two changes that masking makes together. It does not freeze the whole block or the propagated feature maps: shared parameters and later representations can still change through other supported paths.

The new native-local-prefix control gives a +10.700 pp Cut acquisition effect and a +14.100 pp cut interaction. Both paired intervals prespecified in this follow-up are above zero. The earlier auxiliary-prefix interaction is -2.280 pp.

Table R2. Normalization support and acquisition

MethodUnsupported denominatorDirect unsupported derivativeArriving (%)
M0Excluded046.40 ± 1.61
MatchedIncluded> 017.18 ± 2.99
AnchorIncluded044.84 ± 2.49

DeeperForward on CIFAR-100, two ten-class tasks, M = 80. Values are mean ± sample SD over five seeds. Anchor gains 27.66 percentage points over Matched. Sources: Section 3, Table 2 and Appendix Table 5.

Changes made: Reframed the contribution around same-state support separation and added the native-local-prefix test alongside the earlier controls (Introduction; Sections 2–3, 5.6–5.7; Appendix I.8).


AuthorsOpenReview ↗

Response 2: W2, W3, Q1 — Denominator design and related work

W2 | Why retain denominator coordinates?

M0 has higher observed new-task means in all three comparisons. Original two-task current accuracy is 44.84% for Anchor and 46.40% for M0. The common-backbone two-task values are 43.60% and 44.72%. Full-pool mean arriving accuracy is 71.033% and 71.858%, respectively. These comparisons measure acquisition separately from final accuracy. Anchor tests whether acquisition can recover while unsupported coordinates remain in numerical normalization. Relative to Matched, the original and full-pool gains are 27.66 and 18.451 percentage points. Removing direct unsupported derivatives therefore recovers much of the acquisition loss while retaining denominator coordinates.

Retaining the denominator is therefore an experimental constraint with a specific purpose: it rules out denominator removal as a necessary explanation of the recovery. If the only comparison were M0 versus Matched, improved acquisition could follow either from removing unsupported derivatives or from changing normalization and supported gradients. Anchor supplies the missing comparison. The same-state equality makes that intervention precise, while the experiments measure the subsequent training consequence.

Changes made: Reported Anchor and M0 acquisition means separately from final accuracy and clarified the diagnostic role of denominator retention (Sections 5.3, 5.6–5.7 and 6).


W3 | Prior work and complete baselines

The revised related-work discussion separates three connections: plasticity loss in BP, negative supervision or asymmetric competition in class-incremental learning, and the forward reuse of locally supervised feature maps. Lyle et al. and Dohare et al.'s Nature study concern broader plasticity; ER-ACE, FOO-VB, and SS-IL provide relevant continual-classifier alternatives. Their full algorithms combine additional objectives, replay, or inference choices. The revised comparison identifies which learning objectives and training rules each control actually tests.

Changes made: Expanded related work on plasticity and continual classifiers, distinguishing full algorithms from the implemented derivative controls (Sections 6–7).


Q1 | Questions referring to weaknesses

W1–W3 address the Questions field's reference to the weaknesses. Together they identify the same-state intervention and retain M0 as the practical comparison. The full-pool and BP controls add evidence to that clarified contribution.

Changes made: Linked the Questions field to W1–W3; the corresponding revisions are in Sections 2–3, 5.6 and 6–7, and Table 3.


Closing

Thank you again for your time and careful consideration. We would value your further feedback on the revised explanation. In particular, any guidance on making the contribution and its relationship to prior work clearer would be appreciated.

Reviewer_1Ko7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors study the forward-forward algorithms in class-incremental continual learning tasks. They identify the problem of gradient propagation through unsupported class output (outputs that have not seen corresponding positive inputs). The problem is even more important for layerwise local algorithms where gradients through unsupported output changes representation used by subsequent layers. By stopping the gradient through unsupported output, they propose Anchor-Forward algorithms, which improves the arriving accuracy on class-incremental learning.

Critical strengths

The setup and problem is well defined. The authors identify the limitation in directly applying Foward-Foward algorithms to class incremental tasks. Solving the limitation leads to arriving accuracy improvements.

Critical weaknesses
  1. Performance improvement is based on mean arriving accuracy. The benefit of Anchor-FF only appears when many classes have not been seen (supported by Figure 5). However, in continual learning, remembering past learned tasks is also important. Unfortunately, all evaluated methods suffer from significant forgetting (figure S6), and Anchor-FF does not improve final accuracy.

  2. It is also not clear to me why studying plasticity tax of unsupported output in FF algorithms. Except for the investigation on depth exposure (section 5.2), it seems that the problem is not specific to local learning algorithms, but the nature of using supervised cross-entropy classifiers. In end-to-end training with backpropagation or local learning with auxiliary classifiers (e.g. the Nokland & Eidnes, 2019 paper mentioned in the section 7), would stopping the gradient through unsupported output also improve the performance? If the proposed gradient-stopping method works beyond FF, then the paper would have a broader impact on class-incremental learning.

Questions for authors
  1. How does FF method compare (under the same architecture) to backpropagation or other previously published layer-local learning methods, both in terms of performance and memory efficiency?
  2. If gradient is not propagated through unsupported output g_u, can one initialize g_u to low values and exclude it from the loss (equation 4)?
4 comments
AuthorsOpenReview ↗

Response 1: W1, W2 — Acquisition, retention, and BP controls

Opening

Thank you for your careful review and constructive questions. Your comments helped us distinguish new-class acquisition from retention more clearly. They also motivated stronger comparisons with backpropagation and additional controls on backward connections. We have revised the analysis and presentation accordingly and respond to each point below.

Added theory and experiments

The same-state identity was already in the submission; the revision clarifies its scope rather than introducing a new theorem. Added evidence includes the full-pool ten-task study, a common-backbone system comparison, two within-BP interventions, and a common-prefix gradient-cut control (Sections 5.6–5.7; Appendix I, Tables 18–21). These distinguish acquisition, retention, training rules, and readouts. A fixed 20-continuation follow-up from five native-local prefixes adds a positive conditional cut interaction.

We have updated the supporting materials with the complete code for our method. Experiment configurations and result-recomputation scripts are also included.

Figure changes

Figure 3(b) and Figure S5 separate the depth and phase cohorts. Figure S10 removes the pooled heterogeneous depth estimate and uses the original three-seed current-versus-neither comparison. Added Figure S11 displays five prespecified paired contrasts from the full-pool, auxiliary-BP, and common-prefix studies, with their distinct endpoints and intervals. Figure 7 in main-text Section 5.7 shows the new native-local-prefix seed pairs and its two contrasts prespecified in this follow-up.

W1 | Acquisition and retention

We now report acquisition and retention separately in Section 5.6 and Table 3. The added full-pool experiment separates these outcomes directly. In the five-seed, ten-task CIFAR-100 study, Anchor improves mean arriving accuracy over Matched by 18.451 percentage points (pp), averaging tasks 2–10 within each seed (97.5% paired interval [15.636, 21.266]). Final accuracies are 8.502%, 8.484%, 8.430%, and 8.984% for local M0, Matched, Anchor, and terminal BP-M0. Mean final old-90-class accuracy is below 1% for every system. The revised Abstract and Introduction focus on acquisition recovery; Section 5.6 and Table 3 report retention separately.

Changes made: Added the full-pool acquisition and retention results (Section 5.6, Table 3; Appendix I.7, Figure S11). The Abstract and Introduction now report acquisition gains.


W2 | Generic gradients and BP controls

The non-target cross-entropy derivative is generic. We tested its role with two within-BP interventions and a common-prefix control. Terminal BP gives a current-class Anchor-minus-Matched effect of -0.500 pp. BP through all 17 native auxiliary losses gives -0.520 pp. Both preserve their own trained prefixes, recipes, and readouts across arms. A stricter comparison shares the complete auxiliary-trained first-task state and changes only second-task block-input backward cuts. The interaction is -2.280 pp (95% interval [-4.463352, -0.096648]). Forward feature transmission remains present.

These controls answer different parts of the question. Terminal BP changes the learning system and final-head readout; the native-auxiliary BP test retains all local objectives and the summed native readout while allowing gradients to cross blocks. Within each prefix-conditioned 2×2 test, the two factors are Matched versus Anchor and Cut versus Coupled. Its interaction is (Anchor − Matched) under Cut minus the same difference under Coupled. A positive value means that cuts strengthen the relative Anchor effect under that restored starting state.

The new follow-up restores five native-local M0 first-task states and runs 20 fixed continuations in one common summed-loss engine. Cut and Coupled differ only in block-input detachment, with optimizer moments, replay and input/dropout streams preserved. Both prefix studies use the same constant-rate second-task recipe, which differs from the original native sequential-update experiment. Current accuracy is 26.46/37.16% for Cut Matched/Anchor and 30.92/27.52% for Coupled Matched/Anchor, always predicting among all 20 seen classes. The two primary contrasts were prespecified in this follow-up. The Cut effect is +10.700 pp (97.5% paired interval [5.920104, 15.479896]); the Cut-minus-Coupled interaction is +14.100 pp (97.5% paired interval [6.314577, 21.885423]). All 20 final states were locked before testing. Backward cuts strengthen the relative Anchor acquisition effect under this native-local starting configuration and continuation recipe. All four arms and old/current/final/AIA endpoints are reported.

Changes made: Added both prefix-conditioned cut tests and the within-BP interventions (Section 5.7; Appendix I.2, I.4–I.5 and I.8; Figures 7 and S11).


AuthorsOpenReview ↗

Response 2: Q1, Q2 — Performance, memory, and unsupported outputs

Q1 | Matched-backbone performance and resources

The new comparison uses the same 17-layer backbone, replay budget, and two ten-class CIFAR-100 tasks. Separate validation on a tuning seed selects one of three recipes per learning system; five subsequent seeds provide the results. Final accuracy is 42.06/22.50/41.61/39.83% for local M0/Matched/Anchor/terminal BP-M0. Local Anchor retains more old-class accuracy than BP (39.62% versus 20.14%) but acquires fewer current examples correctly (43.60% versus 59.52%). Local prediction sums 17 goodness scores; terminal BP uses its final head, so this is a system comparison. Mean second-task peak allocation is 1,272.42 MiB per local arm versus 3,986.18 MiB for terminal BP. The full-pool results reverse the two-task final ordering against BP. Native auxiliary BP-M0 reaches 47.45% final accuracy with a 17-loss sum and 17-score readout. Its mean peak allocation is 4,113.986 MiB. Its delayed summed-backward updates differ from sequential native local updates.

Table R1. Accuracy and memory on the 17-layer backbone

SystemOld (%)Current (%)Final (%)Peak (MiB)
Local M039.4044.7242.061,272.42
Local Matched27.9017.1022.501,272.42
Local Anchor39.6243.6041.611,272.42
Terminal BP-M020.1459.5239.833,986.18
Auxiliary BP-M032.0262.8847.454,113.99

CIFAR-100, two ten-class tasks. Accuracies are means over five seeds; memory is mean peak allocation in task 2. Sources: Appendix I.1 and I.4, Tables 18 and 20.

Changes made: Added the common-backbone accuracy and resource comparisons, including the auxiliary BP-M0 system and its update/readout rules (Section 5.7; Appendix I.1 and I.4).


Q2 | Low initialization with exclusion

M0 already excludes unsupported scores from the loss while retaining feature capacity. We distinguish this baseline from your combined proposal of low initialization and exclusion. For fixed logits, sending an unsupported loss-side offset to minus infinity approaches the M0 loss and derivatives. Class-associated feature maps also feed later blocks, so changing their initial parameters can change forward representations.

Changes made: Added the loss-side offset limit (Section 6; Appendix I.6).


Closing

Thank you again for your time and suggestions. We would be grateful for your further thoughts on these results. Please let us know if any point needs clarification or closer analysis.

Reviewer_1Ko7OpenReview ↗

Further discussions about weakness and questions

I thank the authors for their effort. The answer to Weakness 1 seems to confirm the limitation that overall performance does not benefit from Anchor. As a result, I would still consider the improvement on arriving accuracy not impactful, unless the authors provide strong arguments on why it is important to focus on arrival accuracy.

Based on my interpretation of the answer to Weakness 2, the benefit of anchor is restricted to specific FF algorithms. If it is the case, could authors elaborate on the motivation to study FF on continual learning?

For the replies to Q1 and Q2, I appreciate the additional experiments and clarifications from the authors. For the two-task study, FF performs better than terminal BP in terms of retention and memory. This could be an interesting perspective to motivate FF, but more extensive studies on 10 tasks and more architectures would be needed.

Overall, I would prefer to keep my initial score unless the authors want further discussions. And I thank the authors again for their efforts.

AuthorsOpenReview ↗

Follow-up to W1, W2 and Q1–Q2

Thank you for following up. We agree with your reading of the 512 KiB experiment. Anchor improved arrival accuracy, while final accuracy was essentially unchanged.

W1 — Why does arrival accuracy matter, and does the gain last?

A low final score alone cannot tell us whether a class was learned and later forgotten. Arrival accuracy measures whether the model can recognize a new class when first trained on it, among all classes seen so far.

To test what lasts, we ran DeeperForward on full-pool CIFAR-100 for ten tasks with a 4 MiB replay cache. AIA averages accuracy over all seen classes at each task boundary. All values are percentages.

MethodAIAFinal, all 100 classesFinal, earlier 90 classes
Matched22.9112.495.91
Anchor34.8914.127.21
Seen-only M035.6114.457.49
Terminal BP38.5318.1810.82

Anchor gained 11.98 points in AIA over Matched. At task ten, it also gained 1.63 points overall and 1.30 on the earlier 90 classes. All three gains appeared in every seed. We added two seeds after observing the first result, keeping the protocol and class order fixed.


W2 and Q1–Q2 — Why study FF in continual learning?

Local learning is worth studying when training memory is limited. Across tasks 2–10, mean peak allocated GPU memory was 1,901 MiB for Anchor versus 4,550 MiB for terminal BP. BP achieved the highest accuracy. Within DeeperForward, blocking direct gradients to unseen classes closed much of the gap between Matched and seen-only M0.

We also extended HCL-FF to ten tasks. Matched and Anchor reached 23.41% and 23.42% AIA, with final accuracies of 8.21% and 7.65%. The two-task gain did not carry over to this second local objective.

We respect your assessment and appreciate the time you have spent helping us improve the paper. Your questions led us to test what survives to task ten, and your suggestion to study other architectures will guide our next experiments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.